WinoGrande: leaderboard
Large-scale Winograd Schema Challenge with 44K problems testing commonsense reasoning through pronoun resolution. Models must determine what a pronoun refers to using world knowledge.
Metric: Accuracy (%). Source: huggingface.co. Source: epoch.ai. Saturation forecast: Estimated already saturated. 80 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Median Human | 94 | |
| 2 | Llama 3.1 405B | 89.2 | #466 |
| 3 | Claude 3 Opus (20240229) | 88.5 | #412 |
| 4 | GPT-4 (0314) | 87.5 | #574 |
| 5 | DeepSeek V2 | 86.3 | #549 |
| 6 | DeepSeek V3 | 85.2 | #321 |
| 7 | Llama 3 70B | 83.5 | #631 |
| 8 | Qwen 2.5 72B | 82.3 | #499 |
| 9 | Llama 3.1 405B Instruct | 82.2 | #457 |
| 10 | GPT-3.5 Turbo (0613) | 81.6 | #1118 |
| 11 | text-davinci-002 | 81.6 | #1225 |
| 12 | Phi-3-medium-128k-instruct | 81.5 | #1013 |
| 13 | Phi-3-small-8k-instruct | 81.5 | #967 |
| 14 | Qwen 2.5 Coder 32B | 80.8 | #667 |
| 15 | Llama 2 70B Base | 80.2 | #1025 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=winogrande · How It Works · Data refreshed daily, snapshot 2026-10-20.